Abstract & Overview
Soil temperature is an important meteorological parameter that affects various physical and chemical reactions in the soil. However, measuring soil temperature is very expensive. Therefore, an attempt to model soil temperature using available agro-climatic variables is very useful. This study aims at finding a better approach for forecasting morning and evening soil temperature at depths 5 cm and 10 cm with the minimum usage of historical soil temperature data. Weekly data collected at Bathalagoda Rice Research and Development Institute for 20 years were used in this study. Commonly used statistical models such as probabilistic and Markov chain models have limitation as they require historical soil temperature in order to obtain forecasts. Hence, Nonlinear Auto Regressive neural network with exogenous input (NARX) models and Feedforward Neural Network (FNN) models were employed to obtain targeted forecasts. Potential input variables for network models were selected based on serial cross correlation and autocorrelation analysis. Models with different combinations of input variables were tested and the best sets of input variables were selected based on the prediction accuracy. The Mean Square Error (MSE) and the coefficient of determination (R2) of the model of predicted values on actual values of test cases were employed to compare the performances of various neural networks. Suitable models with relatively simple structures were selected and their robustness was analyzed. NARX models showed lower MSE values and higher R2 values than FNN models, suggesting that they could be used as more reliable and suitable models to predict weekly mean soil temperatures at depths 5 cm and 10 cm in Bathalagoda area. Sri Lankan Journal of Applied Statistics, Volume 13 (2012), p. 39-59 DOI: http://dx.doi.org/10.4038/sljastats.v13i0.5123